Google open-sources TorchTPU, a native PyTorch back end for TPUs
- ID
- 37846
- Status
- summarized
- Published
- 21 Oct 2026, 12:40 AM
- Fetched
- 21 Oct 2026, 8:34 PM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://github.com/google-pytorch/torch_tpu
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.0
- Created
- 21 Oct 2026, 8:35 PM
- Tags
- Audience
- developersai_ml_learners
What happened
Google's google-pytorch org has published torch_tpu, an open-source PyTorch backend that runs PyTorch models on TPUs via custom ATen kernels, an XLA compilation cache, torch.compile() graph-mode integration, and distributed training support (data and tensor parallelism). The repo ships examples for Llama, Qwen, ResNet, DLRM, minGPT and UNet, plus Marimo notebooks that the README says are easiest to run on a v6e GCE instance. It has ~502 stars, 25 forks and 4,470 commits, and drew 354 points / 74 comments on Hacker News.
Why it matters
If your training stack is PyTorch and you have (or can get) GCP TPU quota, this is the first-party path to keep your model code and torch.compile() workflow instead of rewriting in JAX — check the ATen kernel coverage for your ops and the compile-cache behaviour before assuming parity, and budget for v6e quota rather than assuming a drop-in pip install.
Discussion angle
Does a native PyTorch TPU backend actually erode JAX's grip on TPUs, or is the real blocker still v6e quota and kernel coverage? Worth walking through which of the listed example models (Llama, Qwen, DLRM) you'd realistically port, and what the 74-comment HN thread flags as missing.